Also, switch from returning false to throwing errors in most verify*()
functions, as well as in `await signatures[*].verified`, in order to be
able to show more informative error messages.
(When config.allow_unauthenticated_stream is set or the message is
AEAD-encrypted.)
The issue was that, when hashing the data for verification, we would
only start hashing at the very end (and keep the message in memory)
because nobody was "pulling" the stream containing the hash yet, so
backpressure was keeping the data from being hashed.
Note that, of the two patches in this commit, only the onePassSig.hashed
property actually mattered, for some reason. Also, the minimum
highWaterMark of 1 should have pulled the hashed stream anyway, I think.
I'm not sure why that didn't happen.
Both those with a 2-byte hash (instead of SHA1 or an AEAD authentication
tag) and those without an S2K specifier (i.e., using MD5 for S2K) -
support for the latter was already broken.
Vulnerabilities can arise not just from generating keys like this, but
from using them as well (if an attacker can tamper with them), hence why
we're removing support.
This also has the effect that we only throw on them when trying to use
the key, instead of when parsing it, and that we don't throw when the
authorized revocation key is specified in a separate direct-key
signature instead of a User ID self-signature (the spec only specifies
including it in a direct-key signature, so that means that we
effectively don't reject them anymore. This is because users that
wanted to use the key, could remove this separate signature, anyway.)
Also, when generating RSA keys in JS, generate them with p < q, as per
the spec.
Also, when generating RSA keys using Web Crypto or Node crypto, swap the
generated p and q around, so that will satisfy p < q in most browsers
(but not old Microsoft Edge, 50% of the time) and so that we can use the
generated u coefficient (p^-1 mod q in OpenPGP, q^-1 mod p in RFC3447).
Then, when signing and verifying, swap p and q again, so that the key
hopefully satisfies Safari's requirement that p > q, and so that we can
keep using u again.
This PR adds four config options to configure whether and how to load
indutny/elliptic: use_indutny_elliptic, external_indutny_elliptic,
indutny_elliptic_path and indutny_elliptic_fetch_options.
Also:
- Use tweetnacl.js instead of indutny/elliptic for curve25519 key generation
- Don't initialize indutny's curve25519, improving performance when using that curve
- Verify NIST signatures using Web Crypto instead of indutny/elliptic when not streaming
- Move KeyPair.sign/verify to ecdsa.js
- Move KeyPair.derive to ecdh.js
- Move keyFromPrivate and keyFromPublic to a new indutnyKey.js file
Keep supporting the old names as well though in `openpgp.generateKey`
and `getAlgorithmInfo`, but not in `openpgp.key.generate` (as it is
recommended that developers use `openpgp.generateKey` instead, and
it now throws when using `numBits` instead of `rsaBits`, so there's
no risk of silent key security downgrade).
The old names are now deprecated, and might be removed in v5.
Don't keep the entire message in memory.
This also fixes an unhandled promise rejection when the input
stream contains an error (e.g. an armor checksum mismatch).
Previously the signature parsing function ignored critical bit on
notations.
This change checks for notations that are marked "critical" but are not
on the known notations list (controlled by config array
`openpgp.config.known_notations`) and triggers parse error if such
a notation have been encountered.
See: #897.
Previous implementation used an object to hold notations so if multiple
notations had the same key name only the last one was visible.
After this change notations are exposed as an array of key-value pairs
that can be converted to a map through `new Map(notations)`.
See #897.
